Fig. 2: Schematic of different representational geometries for 4 conditions in the neural activity space and their properties. | Nature Communications

Fig. 2: Schematic of different representational geometries for 4 conditions in the neural activity space and their properties.

From: Neural representational geometries reflect behavioral differences in monkeys and recurrent neural networks

Fig. 2

A Left: factorized or disentangled representations where the 4 points are arranged on a square. The shape (circle vs triangle) and color (red vs blue) are encoded along two orthogonal directions. This geometry supports the representation of shape (and color) in abstract format, i.e., high CCGP. Right: Random representation where the 4 points are placed at random locations in the activity space. This geometry does not support the representation of the shape in abstract format, i.e., low CCGP. B Left: Low shattering dimensionality, where the 4 points are placed at the vertices of a square. The shattering dimensionality is low because not all the dichotomies can be decoded by a linear decoder due to the XOR configuration (purple and green circles). Right: High shattering dimensionality supports the decoding of a higher number of dichotomies, including the one not linearly decodable, i.e., XOR.

Back to article page